Verlag Harri Deutsch, Thun - Frankfurt am Main, Reihe Physik 53 (PhD thesis), ISBN 3-8171-1481-8, 110 pages (1995) (bibtex, paper.pdf, paper.ps.gz)

Labeled graphs and dynamic link matching for face recognition and scene analysis.

Laurenz Wiskott


Abstract:

In many neural net applications visual data are represented as vectors, although it is known that this form of representation lacks syntactical structure. Labeled graphs have been proposed as a data format which provides the missing relational information. The present work argues that labeled graphs of perceptual patterns can be generated and processed based on simple principles. Complex and flexible object representations can be derived from single examples by graph matching.

Dynamic Link Matching has been developed as a biologically motivated neural mechanism for graph matching. This work discusses the principles as well as the advantages and drawbacks of Dynamic Link Matching compared to other neural systems. A complete face recognition system based on Dynamic Link Matching is developed. In contrast to previous systems, the dynamics is autonomous, and matching between graphs of different size is made possible by an attention window. The performance is demonstrated for faces of different perspective or facial expressions against a gallery of 112 neutral frontal views.

For more technical applications, Elastic Graph Matching has been developed as an algorithmic counterpart to Dynamic Link Matching. In this work the system is developed further in several aspects: object adapted graphs allow comparisons between very different views, efficiency has been increased significantly by separating graph generation from recognition, and phase information of the Gabor transform is used to increase matching accuracy. The key role is played by a newly introduced graph structure, called General Face Knowledge. It is based on a collection of individual sample faces, but it also represents faces that can be obtaind by combining subparts of the samples. By this means, new faces can be processed without having a reference model of the individual person. Recognition results on galleries of 300 faces are presented.

The determination of facial attributes serves as a second demonstration. The General Face Knowledge can be used to generate composite or phantom faces very similar to the original. If facial attributes such as gender or the presence of a beard are known for the sample faces of the General Face Knowledge, these attributes can be transfered to the phantom face. On that basis the facial attributes of the original face can be determined in a very simple and general way.

And finally, Elastic Graph Matching is applied to the recognition of occluded objects in cluttered scenes. Two different algorithms are presented. The first allows recognition of known objects between and behind unknown distractors. The second one requires that all objects in the scene are known to the system. It processes the scene from front to back and in addition determines the order of the objects in depth.

Contents:

1. Introduction
2. Labeled Graphs for Object Representation
3. Dynamic Link Matching
4. Face Recognition by Dynamic Link Matching
5. Face Recognition by Elastic Graph Matching
6. Phantom Faces and Face Analysis
7. Recognizing Objects in Cluttered Scenes
8. Conclusion
A. Preprocessing with Gabor Wavelets
B. Zusammenfassung in deutscher Sprache


Relevant Projects:


Juli 7, 2003, Laurenz Wiskott, http://www.neuroinformatik.ruhr-uni-bochum.de/PEOPLE/wiskott/